US20190229992A1 - System and Methods for Auto-Tuning Big Data Workloads on Cloud Platforms - Google Patents
System and Methods for Auto-Tuning Big Data Workloads on Cloud Platforms Download PDFInfo
- Publication number
- US20190229992A1 US20190229992A1 US16/102,852 US201816102852A US2019229992A1 US 20190229992 A1 US20190229992 A1 US 20190229992A1 US 201816102852 A US201816102852 A US 201816102852A US 2019229992 A1 US2019229992 A1 US 2019229992A1
- Authority
- US
- United States
- Prior art keywords
- data
- cloud
- tuner
- data engine
- cloud platform
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 43
- 238000004891 communication Methods 0.000 claims abstract description 19
- 230000003068 static effect Effects 0.000 claims abstract description 17
- 238000013500 data storage Methods 0.000 claims abstract description 9
- 230000003116 impacting effect Effects 0.000 claims abstract description 3
- 238000012545 processing Methods 0.000 claims description 8
- 238000010801 machine learning Methods 0.000 claims description 7
- 238000013499 data model Methods 0.000 claims description 6
- 238000005192 partition Methods 0.000 claims description 4
- 239000000203 mixture Substances 0.000 claims description 3
- 238000012986 modification Methods 0.000 claims description 3
- 230000004048 modification Effects 0.000 claims description 3
- 238000003339 best practice Methods 0.000 claims description 2
- 230000009467 reduction Effects 0.000 description 18
- 238000013459 approach Methods 0.000 description 11
- 238000005457 optimization Methods 0.000 description 9
- 230000008569 process Effects 0.000 description 8
- 230000006870 function Effects 0.000 description 7
- 239000003638 chemical reducing agent Substances 0.000 description 3
- 238000010276 construction Methods 0.000 description 3
- 238000009434 installation Methods 0.000 description 3
- 238000013178 mathematical model Methods 0.000 description 3
- 230000009291 secondary effect Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000002596 correlated effect Effects 0.000 description 2
- 238000007726 management method Methods 0.000 description 2
- 230000003252 repetitive effect Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000003203 everyday effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- ZLIBICFPKPWGIZ-UHFFFAOYSA-N pyrimethanil Chemical compound CC1=CC(C)=NC(NC=2C=CC=CC=2)=N1 ZLIBICFPKPWGIZ-UHFFFAOYSA-N 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 238000013341 scale-up Methods 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/21—Design, administration or maintenance of databases
- G06F16/217—Database tuning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/27—Replication, distribution or synchronisation of data between databases or within a distributed database system; Distributed database system architectures therefor
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/08—Configuration management of networks or network elements
- H04L41/0803—Configuration setting
- H04L41/0823—Configuration setting characterised by the purposes of a change of settings, e.g. optimising configuration for enhancing reliability
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/08—Configuration management of networks or network elements
- H04L41/0876—Aspects of the degree of configuration automation
- H04L41/0886—Fully automatic configuration
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/12—Discovery or management of network topologies
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/50—Network service management, e.g. ensuring proper service fulfilment according to agreements
- H04L41/5003—Managing SLA; Interaction between SLA and QoS
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/01—Protocols
- H04L67/10—Protocols in which an application is distributed across nodes in the network
- H04L67/1001—Protocols in which an application is distributed across nodes in the network for accessing one among a plurality of replicated servers
-
- H04L67/1002—
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/01—Protocols
- H04L67/10—Protocols in which an application is distributed across nodes in the network
- H04L67/1097—Protocols in which an application is distributed across nodes in the network for distributed storage of data in networks, e.g. transport arrangements for network file system [NFS], storage area networks [SAN] or network attached storage [NAS]
Definitions
- the preset invention is directed to systems and methods that may provide a common abstraction layer to tune various modules of a big data engine, data set and workloads on cloud platforms. More specifically, the present invention may automatically tune such modules to improve cost, decrease cost or both.
- Big Data generally refers to a collection of data sets that may be difficult to process using traditional data processing applications due to its size and/or complexity. Due to its size, processes to capture, process, manage, and/or curate the data within a reasonable time may be quite difficult. Big Data has been defined to include a high volume of data, high velocity processing speed, high variety of information, and high veracity or accuracy of information and analytics.
- Big Data may be utilized in virtually any service or product. Ranging from science and research applications (such as, for example, data collected by the Large Hadron Collider, or information related to the human genome) to those in everyday life (such as, for example, services provided by Facebook, Amazon, eBay), the ability to process Big Data and determine meaningful conclusions therefrom is becoming a necessity.
- Big Data has grown, computing resources necessary to efficiently handle such data is required. Due to certain constraints—ranging from cost to physical space—cloud computing has provided resources that can be tapped in order to process Big Data. Big Data systems may be quite powerful and may process petabytes of data and run on upwards of thousands of machines. However, there are periods when such systems are underutilized and other periods when such systems are over-taxed. Accordingly, it is desirable to have a Big Data system that can be adapt to workload and scale up and down, thereby leading to better utilization of resources and reducing costs.
- Cloud computing has what can be referred to as an elastic property, in that virtual machines in the cloud can be used on-demand on an as-needed basis, thereby providing resources that are suitable for specific processing tasks or functions.
- the flexibility of resources in a cloud computing environment can be difficult to control. Determining the proper amount of resources at the outset of an activity can be difficult, particularly when the loads of certain applications fluctuate.
- a large-scale cloud computing environment the use of various processing and computing resources may be available at a certain cost rate. There is often a balance between the cost of resources and optimal performance. In some situations, cloud computing products or services may be offered to provide better financial rates, but with certain drawbacks. Accordingly, a large-scale cloud processor may seek to either rectify or mitigate drawbacks to obtain optimal performance at desirable rates.
- execution engines such as Map Reduce (M/R), Tez, Presto, and/or Spark may provide a set of knobs or configuration parameters that control the behavior of the execution engine. Accordingly, it is desirable to determine a good set of parameters for SQL workloads amongst different engines. However, it may be tricky to identify an optimal, or even a good, set of parameters for a specific workload. The list of parameters is long and many of the parameters are correlated in unexpected ways. For example, in M/R, mapper memory and input split size are correlated since a good value for the memory parameter may depend on the split size.
- an “Extract, Transform, and Load” (ETL) engineer will determine a set of parameters after analyzing a few important workloads. However, these parameters may not be optimal for all workloads. Moreover, as the queries and data change, parameters determined at one point in time may not be optimal over time. An automated approach that can recommend an optimal set of configuration values for each workload is desirable and may present a scalable option.
- an auto-tuning framework is desirable.
- an auto-tuning system framework may include one or more attributes of: (i) collecting relevant metrics from the data engine; collecting relevant statistics from the data sets; (iii) collecting relevant workload properties; (iv) recommending changes that may use rule-based, cost-based, and/or machine learning based algorithms to generate tuning recommendations; (v) applying recommendations to the data engine; (vi) applying recommendations to the data set; (vii) applying recommendations to the workload; and/or (viii) receiving information from the cloud platform.
- the present invention provides such desirable attributes.
- aspects may include a system for automatically tuning big data workloads across various cloud platforms, the system being in communication with a cloud platform and a user, the cloud platform including data storage and a data engine, the system comprising: a system information module in communication with the cloud platform: a static tuner in communication with the system information module; a cloud tuner in communication with the static tuner and the user; and an automation module in communication with the cloud tuner.
- aspects may include a method for automatically tuning big data workloads across various cloud platforms, the method being applied across different cloud platforms, data models, and data storage formats, the method utilizing a tuning module in communication with a cloud platform and a user, the cloud platform including data storage and a data engine, the method comprising: extracting information from the cloud platform, the information impacting or associated with the performance of the big data workload; determining recommendations based at least in part on the information extracted from the cloud platform; iterating through different valid hardware configurations to determine desired or optimal hardware and data engine configuration; and applying the determined desired or optimal hardware and data engine configuration to the data engine.
- FIG. 1 illustrates an exemplary arrangement of system components for auto-tuning big data workloads on cloud platforms, in accordance with some embodiments of the present invention.
- FIG. 2 depicts an exemplary iterative algorithm for determining optimal values for certain parameters, in accordance with some embodiments of the present invention.
- FIG. 3 illustrates an exemplary chart, showing reduction of query cost, in accordance with some embodiments of the present invention.
- FIG. 4 illustrates an exemplary chart, showing predicted reduction in cost compared with observed reduction in cost, in accordance with some embodiments of the present invention.
- the present invention is directed to systems and methods that may provide a common abstraction layer to tune various modules of a big data engine, data set and workloads on cloud platforms. More specifically, the present invention may automatically tune such modules to improve cost, decrease cost or both.
- some embodiments of the present invention may determine a good or optimal set of parameters for SQL workloads amongst different engines.
- Some embodiments of the present invention may provide scalable systems and methods that may recommend an optimal set of configuration values for each workload.
- some embodiments of the present invention may comprise an auto-tuning system framework that may include one or more attributes of: (i) collecting relevant metrics from the data engine; collecting relevant statistics from the data sets; (iii) collecting relevant workload properties; (iv) recommending changes that may use rule-based, cost-based, and/or machine learning based algorithms to generate tuning recommendations; (v) applying recommendations to the data engine; (vi) applying recommendations to the data set; (vii) applying recommendations to the workload; and/or (viii) receiving information from the cloud platform.
- an auto-tuning system framework may include one or more attributes of: (i) collecting relevant metrics from the data engine; collecting relevant statistics from the data sets; (iii) collecting relevant workload properties; (iv) recommending changes that may use rule-based, cost-based, and/or machine learning based algorithms to generate tuning recommendations; (v) applying recommendations to the data engine; (vi) applying recommendations to the data set; (vii) applying recommendations to the workload; and/or (viii) receiving information from the cloud platform.
- the present invention is agnostic as to the specific cloud platform, big data engine, and/or workload data.
- recommendations may consider the dynamism of cloud platforms, the specific identity of cloud platforms (AWS, Microsoft Azure, Google Cloud, etc.) is not a determining factor and other cloud platforms not discussed may be used by the present invention.
- big data engines such as Apache Hive, Apache Tez, Apache Spark, and Presto are discussed, it is contemplated that systems and methods of the present invention are not limited to specific data engines, and instead can be utilized with various big data engines either not listed here, or not yet in use.
- Workloads discussed may include ETL, ad-hoc, reporting, data science, and AI data, other workloads and datasets may be used.
- Recommendations of auto-tuning may include number of machines, machine type and/or the machine life cycle policy.
- machine life cycle policies may include, but are not limited to, preemptible, spot-nodes, spot blocks, on-demand.
- System 100 generally comprises three (3) elements: cloud platform 110 , user 120 , and auto-tuning module 130 .
- Cloud platform 110 may comprise a data engine 111 and a data storage 112 .
- the user 120 may interact with the data engine 111 to, for example, send query requests 121 to the data engine 111 , and receive results 122 from the data engine 111 .
- Auto-tuning module 130 may comprise a system information submodule 131 , static tuner 132 , a cloud tuner 133 , and an automation sub-module 134 .
- the system information sub-module 131 may receive information from the cloud platform 110 and send information to the static tuner 132 .
- Static tuner 132 may be in communication with cloud tuner 153 , which may also receive inputs 123 from user 120 .
- Cloud tuner 133 may send information to the automation sub-module 134 , which in turn may convey information back to the cloud platform 110 .
- system information sub-module 131 may gather information such as, but not limited to: SQL queries; machine learning jobs; data processing jobs written in Map-Reduce, Scala, Pig etc.; metadata, such as table and/or column data; system metrics, for example associated with CPUs, networks, etc.; data engine metrics, such as the number of tasks pending or processing, spills, attempts, etc.; cloud hardware information, such as machine configuration, billing policies, etc.; inventory of compute resources provided by the cloud platform like machine types; price per unit of compute resources; and/or policies to control budget and service level agreements (SLAs) from administrators.
- SQL queries such as, but not limited to: SQL queries; machine learning jobs; data processing jobs written in Map-Reduce, Scala, Pig etc.
- metadata such as table and/or column data
- system metrics for example associated with CPUs, networks, etc.
- data engine metrics such as the number of tasks pending or processing, spills, attempts, etc.
- cloud hardware information such as machine configuration, billing policies, etc.
- Each of the cloud platforms, engines, data sets and workloads may provide an interface to extract this information.
- a data warehouse may be used to store this information.
- the data warehouse may support both real time as well as batch analysis.
- the static tuner 132 may use data captured by the system information sub-module 131 .
- the input and recommendation may depend at least in part on the sub-system of the data engine.
- the input may be queries and metadata.
- Recommendations may then be generated using one or more of: (i) heuristics or known industry best practices (for example, data partitioning); (ii) cost-based (for example, using sort or bucket data); and/or (iii) machine learning techniques, which may for example be used to find, identify, or determine optimal engine settings.
- Algorithms used by the static tuner 132 may utilize historical data of usage, metrics and configuration of the big data engine captured and stored in the system information sub-module.
- the cloud tuner 133 may use the static tuner 132 and policies to iterate through valid hardware configurations and determine desired or optimal hardware and data engine configuration.
- the static tuner 132 and cloud tuner 133 may also manage policies for current recommendations. For example, if materialized views are recommended and there is a storage or freshness of data constraint, then the materialized view may be refreshed as per policy.
- the cloud tuner 133 may implement algorithms to tune the database by considering factors that may include, but are not limited to, the following: (i) size of cluster, (ii) machine types for one or more clusters; (iii) storage policies and cost; and/or (iv) network topology of one or more clusters.
- the system may utilize heuristics to reduce the number of hardware configurations considered by the algorithms. These heuristics may be defined as policies provided by experts in the field and may be used as another input to the system. Some heuristics in accordance with some embodiments of the present invention are discussed below.
- the automation submodule 134 may apply the recommendation on the data engine 111 . This may be accomplished automatically or through direct or indirect user instruction.
- the automation submodule may desire or require changes to current engines 111 to accept a new hardware configuration and/or use the application programming interface (API) of a cloud platform to modify or alter hardware on a live data engine installation with acceptable, little, or no down time.
- API application programming interface
- system 10 may further comprise a module to recognize queries failing due to run time errors and implementing a fix to prevent this error. Similarly, if a query is repetitive, the performance or cost of the query may be iteratively improved.
- optimization techniques As noted above, there are various optimization techniques that may be used in auto-tuning algorithms. Such techniques may be applied to various modules in a data engine, including an overall data model, execution engine, clusters, and/or storage. Optimization in each of these modules is discussed in turn.
- data operations may be optimized by storing data using specific data structures.
- this feature is typically called materialized views.
- materialized views may be created, available storage is often a limiting factor or main consideration.
- considerations may turn to the cost of storage.
- the present invention may provide operational support to keep cost under control as per policies determined by the administrator.
- the data model recommended by some embodiments of the present invention may use queries run and metadata to provide following recommendations: (i) table partitions; (ii) sort columns; (iii) denormalized or pre-joined tables; and/or (iv) narrow tables.
- execution engines in typical installations machine and resource configuration may be substantially or practically static. However, in dynamic environments the number of machines and resources (such as but not limited to CPU, memory, network, disk, etc.) may vary. Accordingly, a similar or even same job across different installations or environments may require different configurations.
- the execution engine recommender may recommend: (i) machine type; (ii) number of nodes; and/or (iii) configuration parameters, such as split size, number of reducers, container size, etc.
- Cluster management may determine the composition in terms of number of machines and machine types. Factors that may be considered may include workload and billing or cost policies. In accordance with some embodiments of the present invention, the cluster management recommender may recommend: (i) spot node policy; and/or (ii) heterogeneous machine composition.
- the storage layers available are object store, SAN, local disk and memory. However, the data to store each layer of storage may be difficult to control and determine.
- the storage recommender may determine which tables, columns, and/or partitions will reside in various layers.
- optimization Approaches/Methodologies In general, an optimization function should be a function of dollar cost and throughput. It is observed that administrators often focus on only one of these functions, often resulting in sub-optimal configurations. For example, execution engine configuration is sub-optimal in the field. In nearly every experiment performed by the inventors, a large fraction of the queries could be optimized by more than sixty percent (60%). This points to the fact that, in most eases, manual efforts to choose optimal configuration fail.
- a model of an execution engine may provide recommendations for SQL workloads without requiring actual execution.
- models may be used in various engines, including but not limited to Map Reduce, Tez, Presto, and Spark, and may be used to automate and provide insights and recommendations for engine configuration.
- Approaches used to determine optimal configurations may include; (i) iterative execution approaches: and (ii) mathematical models.
- jobs may be executed multiple times with different configuration parameters to determine an optimal configuration.
- these approaches focus on techniques to converge towards a good configuration using lesser number of steps.
- a noisy gradient may be used to converge to a solution and apply stochastic optimization to parameter tuning.
- a mathematical model may be used to predict runtime/cost of jobs for a particular configuration.
- the search over the parameter space to find the optimal configuration may be performed using the model without having to actually run the job.
- ⁇ i 1 n ⁇ ⁇ i ⁇ ⁇ i
- n is the number of containers launched
- ⁇ i is the container memory for the i th container launched
- ⁇ i is the execution time for the i th container launched for job.
- the product is a proxy for the cost of running a container.
- the sum is a proxy for the cost of running a query. This represents both memory and CPU resources consumed, and may correlate to both dollar cost and throughput of a big data stack.
- Map Reduce may include, but are not limited to mapreduce.map.memory .mb, mapreduce.reduce.memory.mb. mapreduce.input.fileinputformat.split.maxsize, mapreduce.input.fileinputformat.split.minsize, and/or hive.exec.reducers.bytes.per.reducer.
- Parameters for Spark may include, but are not limited to spark.executor.cores, spark.executor.memory, spark.hadoop.mapreduce.input.fileinputformat.split.maxsize, spark.hadoop.mapreduce.input.fileinputformat.split.minsize, spark.sql.shuffle.partitions, and/or spark.sql.autoBroadcastJoinThreshold.
- FIG. 2 illustrates steps to search optimal values for two parameters.
- Graphical representation 210 shows the search space for two parameters.
- Graphical representation 220 shows discrete values being explored in each axis.
- Graphical representation 230 illustrates the search space being restricted based on domain knowledge.
- graphical representation 240 shows an algorithm reiterating itself through each parameter, choosing an optimal point, and then moving to the next parameter.
- FIG. 3 illustrates a chart 30 , showing data for three different queries ( 311 , 312 , 333 ), depicted against a vertical axis 320 showing percentage of cost reduction.
- the first query 311 showed a cost reduction of approximately 65%, shown at 331 .
- the second query 312 showed a cost reduction of approximately 75%, shown at 332 .
- the third query 313 showed a cost reduction of approximately 15%. Again, each auto-tuned result showed significant improvement and savings over settings chosen by database administrators.
- An execution model was created that substantially replicated an execution engine, based on a reduced set of parameters.
- this model may also utilize various statistics about data sizes and selectivities of various operators as input. Such statistics may be obtained by collecting metrics from a previous run, or by using statistics from a database catalog. Metrics from a previous run may be suitable for ETL or reporting queries, while statistics from a database catalog may be suitable for ad-hoc queries.
- This model outputs the result of the optimization function as described above.
- the graph 40 at FIG. 4 illustrates the benefit predicted by our model and the actual observed benefit for these queries.
- Each query ( 411 , 412 , 413 , 414 ) is set forth along the x axis, while the y axis 420 shows the percentage of cost reduction. It can be seen that the first query 411 showed a predicted cost reduction ( 431 A) of approximately 82%, while an observed cost reduction ( 431 B) was approximately 76%.
- the second query 412 showed a predicted cost reduction ( 432 A) of approximately 80%, while an observed cost reduction ( 432 B) was approximately 82%.
- the third query 413 showed a predicted cost reduction ( 433 A) of approximately 72%, while an observed cost reduction ( 433 B) was approximately 70%.
- the fourth query 414 showed a predicted cost reduction ( 434 A) of approximately 80%, while an observed cost reduction ( 434 B) was approximately 78%. It can be seen that, in general, the actual savings closely match the predicted savings indicating that the model is sufficiently accurate.
- Insight 1 Container Shape Should Match Instance Family Shape. Yarn allocates containers on two dimensions—memory and virtual central processing unit (vcpu). Each container is given 1 vcpu and some memory. The memory/vcpu of the containers should match the memory/vcpu ratio of the machine type. Otherwise resources may be unnecessarily wasted.
- vcpu virtual central processing unit
- Insight 2 Avoid Spills in Tasks. Spills are expensive because each spill leads to an extra write and read of all the data. Spills should be avoided at all costs. In general, spills may be avoided by providing adequate memory to each task.
- Insight 3 Decrease Memory Per Task to Choose Cheaper Instance Type. (On cloud platforms, machines with higher memory/cpu are generally more expensive for the same CPU type. Decrease the memory per task and consequently increase parallelism to choose a cheaper instance type. As long as tasks do not spill, the total work done in terms of IO, CPU and network traffic is independent of the parallelism factor of the tasks. For example, the total data read and processed will be the same if the number of mappers is 100 or 1000. If a job can be tuned to avoid spills on a cheaper instance with same compute but lesser memory than original instance, then it is generally a good idea to move to cheaper instance for saving cost without any performance degradation.
- Insight 4 Beware of Secondary Effects of High Parallelism.
- Parallelism cannot be increased indefinitely.
- JVM Java virtual machine
- Insight 5 In Spark Environments, Prefer Fat Executors. This insight is specific to Spark, where there is an additional parameter of cores per executor. Given a certain number of cores per machine, there may be a choice of either running many executors with fewer cores per executor (thin executors), or fewer executors with more cores per executor (fat executors). It has been observed that fat executors generally provide better performance, due to better memory utilization across cores, reduced number of replicas of broadcast tables, and lesser overheads due to more tasks running in the same executor process.
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Automation & Control Theory (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Stored Programmes (AREA)
Abstract
Description
- This application claims priority to U.S. Provisional Patent Application No. 62/620,543, filed on 23 Jan. 2018, entitled “Systems and Methods for Auto-Tuning Big Data Workloads on Cloud Platforms,” which is incorporated herein by reference in its entirety. In general, the preset invention is directed to systems and methods that may provide a common abstraction layer to tune various modules of a big data engine, data set and workloads on cloud platforms. More specifically, the present invention may automatically tune such modules to improve cost, decrease cost or both.
- The term “Big Data” generally refers to a collection of data sets that may be difficult to process using traditional data processing applications due to its size and/or complexity. Due to its size, processes to capture, process, manage, and/or curate the data within a reasonable time may be quite difficult. Big Data has been defined to include a high volume of data, high velocity processing speed, high variety of information, and high veracity or accuracy of information and analytics.
- Big Data may be utilized in virtually any service or product. Ranging from science and research applications (such as, for example, data collected by the Large Hadron Collider, or information related to the human genome) to those in everyday life (such as, for example, services provided by Facebook, Amazon, eBay), the ability to process Big Data and determine meaningful conclusions therefrom is becoming a necessity.
- As Big Data has grown, computing resources necessary to efficiently handle such data is required. Due to certain constraints—ranging from cost to physical space—cloud computing has provided resources that can be tapped in order to process Big Data. Big Data systems may be quite powerful and may process petabytes of data and run on upwards of thousands of machines. However, there are periods when such systems are underutilized and other periods when such systems are over-taxed. Accordingly, it is desirable to have a Big Data system that can be adapt to workload and scale up and down, thereby leading to better utilization of resources and reducing costs.
- Cloud computing has what can be referred to as an elastic property, in that virtual machines in the cloud can be used on-demand on an as-needed basis, thereby providing resources that are suitable for specific processing tasks or functions. However, the flexibility of resources in a cloud computing environment can be difficult to control. Determining the proper amount of resources at the outset of an activity can be difficult, particularly when the loads of certain applications fluctuate.
- In a large-scale cloud computing environment, the use of various processing and computing resources may be available at a certain cost rate. There is often a balance between the cost of resources and optimal performance. In some situations, cloud computing products or services may be offered to provide better financial rates, but with certain drawbacks. Accordingly, a large-scale cloud processor may seek to either rectify or mitigate drawbacks to obtain optimal performance at desirable rates.
- Due to various reasons (including, but not limited to, costs, availability, efficiency, speed, and/or other attributes) companies may be running Big Data workloads on various cloud platforms. Configuration tuning of these data engines is an important but difficult undertaking. Cloud platforms may further add to this complexity due to elasticity of compute resources and availability of different machine types. Data engineers often have to choose the correct machine type and number of machines along with the other configuration options.
- For example, execution engines such as Map Reduce (M/R), Tez, Presto, and/or Spark may provide a set of knobs or configuration parameters that control the behavior of the execution engine. Accordingly, it is desirable to determine a good set of parameters for SQL workloads amongst different engines. However, it may be tricky to identify an optimal, or even a good, set of parameters for a specific workload. The list of parameters is long and many of the parameters are correlated in unexpected ways. For example, in M/R, mapper memory and input split size are correlated since a good value for the memory parameter may depend on the split size.
- Typically, an “Extract, Transform, and Load” (ETL) engineer will determine a set of parameters after analyzing a few important workloads. However, these parameters may not be optimal for all workloads. Moreover, as the queries and data change, parameters determined at one point in time may not be optimal over time. An automated approach that can recommend an optimal set of configuration values for each workload is desirable and may present a scalable option.
- Accordingly, an auto-tuning framework is desirable. Specifically desirable is an auto-tuning system framework that may include one or more attributes of: (i) collecting relevant metrics from the data engine; collecting relevant statistics from the data sets; (iii) collecting relevant workload properties; (iv) recommending changes that may use rule-based, cost-based, and/or machine learning based algorithms to generate tuning recommendations; (v) applying recommendations to the data engine; (vi) applying recommendations to the data set; (vii) applying recommendations to the workload; and/or (viii) receiving information from the cloud platform. The present invention provides such desirable attributes.
- In accordance with some embodiments of the present invention, aspects may include a system for automatically tuning big data workloads across various cloud platforms, the system being in communication with a cloud platform and a user, the cloud platform including data storage and a data engine, the system comprising: a system information module in communication with the cloud platform: a static tuner in communication with the system information module; a cloud tuner in communication with the static tuner and the user; and an automation module in communication with the cloud tuner.
- In accordance with some embodiments of the present invention, aspects may include a method for automatically tuning big data workloads across various cloud platforms, the method being applied across different cloud platforms, data models, and data storage formats, the method utilizing a tuning module in communication with a cloud platform and a user, the cloud platform including data storage and a data engine, the method comprising: extracting information from the cloud platform, the information impacting or associated with the performance of the big data workload; determining recommendations based at least in part on the information extracted from the cloud platform; iterating through different valid hardware configurations to determine desired or optimal hardware and data engine configuration; and applying the determined desired or optimal hardware and data engine configuration to the data engine.
- These and other aspects will become apparent from the following description of the invention taken in conjunction with the following drawings, although variations and modifications may be affected without departing from the spirit and scope of the novel concepts of the invention
- The present invention can be more fully understood by reading the following detailed description together with the accompanying drawings, in which like reference indicators are used to designate like elements. The accompanying figures depict certain illustrative embodiments and may aid in understanding the following detailed description. Before any embodiment of the invention is explained in detail, it is to be understood that the invention is not limited in its application to the details of construction and the arrangements of components set forth in the following description or illustrated in the drawings. The embodiments depicted are to be understood as exemplary and in no way limiting of the overall scope of the invention. Also, it is to be understood that the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting. The detailed description will make reference to the following figures, in which:
-
FIG. 1 illustrates an exemplary arrangement of system components for auto-tuning big data workloads on cloud platforms, in accordance with some embodiments of the present invention. -
FIG. 2 depicts an exemplary iterative algorithm for determining optimal values for certain parameters, in accordance with some embodiments of the present invention. -
FIG. 3 illustrates an exemplary chart, showing reduction of query cost, in accordance with some embodiments of the present invention. -
FIG. 4 illustrates an exemplary chart, showing predicted reduction in cost compared with observed reduction in cost, in accordance with some embodiments of the present invention. - Before any embodiment of the invention is explained in detail, it is to be understood that the present invention is not limited in its application to the details of construction and the arrangements of components set forth in the following description or illustrated in the drawings. The present invention is capable of other embodiments and of being practiced or being carried out in various ways. Also, it is to be understood that the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting.
- The matters exemplified in this description are provided to assist in a comprehensive understanding of variety exemplary embodiments disclosed with reference to the accompanying figures. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the exemplary embodiments described herein can be made without departing from the spirit and scope of the claimed invention. Descriptions of well-known functions and constructions are omitted for clarity and conciseness. Moreover, as used herein, the singular may be interpreted in the plural, and alternately, any term in the plural may be interpreted to be in the singular.
- The present invention is directed to systems and methods that may provide a common abstraction layer to tune various modules of a big data engine, data set and workloads on cloud platforms. More specifically, the present invention may automatically tune such modules to improve cost, decrease cost or both. In providing such systems and methods, some embodiments of the present invention may determine a good or optimal set of parameters for SQL workloads amongst different engines. Some embodiments of the present invention may provide scalable systems and methods that may recommend an optimal set of configuration values for each workload.
- As noted above, some embodiments of the present invention may comprise an auto-tuning system framework that may include one or more attributes of: (i) collecting relevant metrics from the data engine; collecting relevant statistics from the data sets; (iii) collecting relevant workload properties; (iv) recommending changes that may use rule-based, cost-based, and/or machine learning based algorithms to generate tuning recommendations; (v) applying recommendations to the data engine; (vi) applying recommendations to the data set; (vii) applying recommendations to the workload; and/or (viii) receiving information from the cloud platform.
- Note that the present invention is agnostic as to the specific cloud platform, big data engine, and/or workload data. In other words, while recommendations may consider the dynamism of cloud platforms, the specific identity of cloud platforms (AWS, Microsoft Azure, Google Cloud, etc.) is not a determining factor and other cloud platforms not discussed may be used by the present invention. Similarly, while big data engines such as Apache Hive, Apache Tez, Apache Spark, and Presto are discussed, it is contemplated that systems and methods of the present invention are not limited to specific data engines, and instead can be utilized with various big data engines either not listed here, or not yet in use. The specific attributes of the big data engine—while relevant to the determination of optimal tuning—do not bear directly on the present invention. Workloads discussed may include ETL, ad-hoc, reporting, data science, and AI data, other workloads and datasets may be used.
- Recommendations of auto-tuning may include number of machines, machine type and/or the machine life cycle policy. Examples of machine life cycle policies may include, but are not limited to, preemptible, spot-nodes, spot blocks, on-demand.
- With reference to
FIG. 1 , anexemplary system 100 of auto-tuning big data workloads on cloud platforms will now be discussed.System 100 generally comprises three (3) elements:cloud platform 110, user 120, and auto-tuningmodule 130. -
Cloud platform 110 may comprise adata engine 111 and adata storage 112. The user 120 may interact with thedata engine 111 to, for example, sendquery requests 121 to thedata engine 111, and receiveresults 122 from thedata engine 111. - Auto-
tuning module 130 may comprise asystem information submodule 131,static tuner 132, acloud tuner 133, and anautomation sub-module 134. The system information sub-module 131 may receive information from thecloud platform 110 and send information to thestatic tuner 132.Static tuner 132 may be in communication with cloud tuner 153, which may also receiveinputs 123 from user 120.Cloud tuner 133 may send information to theautomation sub-module 134, which in turn may convey information back to thecloud platform 110. - More specifically, the system information sub-module 131 may gather information such as, but not limited to: SQL queries; machine learning jobs; data processing jobs written in Map-Reduce, Scala, Pig etc.; metadata, such as table and/or column data; system metrics, for example associated with CPUs, networks, etc.; data engine metrics, such as the number of tasks pending or processing, spills, attempts, etc.; cloud hardware information, such as machine configuration, billing policies, etc.; inventory of compute resources provided by the cloud platform like machine types; price per unit of compute resources; and/or policies to control budget and service level agreements (SLAs) from administrators.
- Each of the cloud platforms, engines, data sets and workloads may provide an interface to extract this information. A data warehouse may be used to store this information. The data warehouse may support both real time as well as batch analysis.
- The
static tuner 132 may use data captured by thesystem information sub-module 131. The input and recommendation may depend at least in part on the sub-system of the data engine. For example, for data modeling, the input may be queries and metadata. Recommendations may then be generated using one or more of: (i) heuristics or known industry best practices (for example, data partitioning); (ii) cost-based (for example, using sort or bucket data); and/or (iii) machine learning techniques, which may for example be used to find, identify, or determine optimal engine settings. Algorithms used by thestatic tuner 132 may utilize historical data of usage, metrics and configuration of the big data engine captured and stored in the system information sub-module. - The
cloud tuner 133 may use thestatic tuner 132 and policies to iterate through valid hardware configurations and determine desired or optimal hardware and data engine configuration. Thestatic tuner 132 andcloud tuner 133 may also manage policies for current recommendations. For example, if materialized views are recommended and there is a storage or freshness of data constraint, then the materialized view may be refreshed as per policy. Thecloud tuner 133 may implement algorithms to tune the database by considering factors that may include, but are not limited to, the following: (i) size of cluster, (ii) machine types for one or more clusters; (iii) storage policies and cost; and/or (iv) network topology of one or more clusters. - Since the number of possible hardware configurations is very high, the system may utilize heuristics to reduce the number of hardware configurations considered by the algorithms. These heuristics may be defined as policies provided by experts in the field and may be used as another input to the system. Some heuristics in accordance with some embodiments of the present invention are discussed below.
- The automation submodule 134 may apply the recommendation on the
data engine 111. This may be accomplished automatically or through direct or indirect user instruction. The automation submodule may desire or require changes tocurrent engines 111 to accept a new hardware configuration and/or use the application programming interface (API) of a cloud platform to modify or alter hardware on a live data engine installation with acceptable, little, or no down time. - Note that it is contemplated that systems in accordance with some embodiments of the present invention may include additional components. For example, system 10 may further comprise a module to recognize queries failing due to run time errors and implementing a fix to prevent this error. Similarly, if a query is repetitive, the performance or cost of the query may be iteratively improved.
- Optimization Techniques. As noted above, there are various optimization techniques that may be used in auto-tuning algorithms. Such techniques may be applied to various modules in a data engine, including an overall data model, execution engine, clusters, and/or storage. Optimization in each of these modules is discussed in turn.
- With regard to data models, note that data operations may be optimized by storing data using specific data structures. In database industries, this feature is typically called materialized views. Note that while multiple materialized views can be created, available storage is often a limiting factor or main consideration. When a storage layer can expand to any capacity, considerations may turn to the cost of storage. The present invention may provide operational support to keep cost under control as per policies determined by the administrator. The data model recommended by some embodiments of the present invention may use queries run and metadata to provide following recommendations: (i) table partitions; (ii) sort columns; (iii) denormalized or pre-joined tables; and/or (iv) narrow tables.
- With regard to execution engines, in typical installations machine and resource configuration may be substantially or practically static. However, in dynamic environments the number of machines and resources (such as but not limited to CPU, memory, network, disk, etc.) may vary. Accordingly, a similar or even same job across different installations or environments may require different configurations. The execution engine recommender may recommend: (i) machine type; (ii) number of nodes; and/or (iii) configuration parameters, such as split size, number of reducers, container size, etc.
- Cluster management may determine the composition in terms of number of machines and machine types. Factors that may be considered may include workload and billing or cost policies. In accordance with some embodiments of the present invention, the cluster management recommender may recommend: (i) spot node policy; and/or (ii) heterogeneous machine composition.
- With regard to storage models, note that in public clouds, data tiering may be an important feature (for example, to reduce latency). The storage layers available are object store, SAN, local disk and memory. However, the data to store each layer of storage may be difficult to control and determine. The storage recommender may determine which tables, columns, and/or partitions will reside in various layers.
- Optimization Approaches/Methodologies. In general, an optimization function should be a function of dollar cost and throughput. It is observed that administrators often focus on only one of these functions, often resulting in sub-optimal configurations. For example, execution engine configuration is sub-optimal in the field. In nearly every experiment performed by the inventors, a large fraction of the queries could be optimized by more than sixty percent (60%). This points to the fact that, in most eases, manual efforts to choose optimal configuration fail.
- Moreover, while optimal configuration could be determined by running a workload iteratively with different values, this method is expensive and impractical. However, a model of an execution engine may provide recommendations for SQL workloads without requiring actual execution. As noted above, such models may be used in various engines, including but not limited to Map Reduce, Tez, Presto, and Spark, and may be used to automate and provide insights and recommendations for engine configuration.
- Approaches used to determine optimal configurations may include; (i) iterative execution approaches: and (ii) mathematical models.
- In an iterative execution approach, jobs may be executed multiple times with different configuration parameters to determine an optimal configuration. As the parameter space is huge, these approaches focus on techniques to converge towards a good configuration using lesser number of steps. For example, a noisy gradient may be used to converge to a solution and apply stochastic optimization to parameter tuning.
- In a mathematical model approach, a mathematical model may be used to predict runtime/cost of jobs for a particular configuration. The search over the parameter space to find the optimal configuration may be performed using the model without having to actually run the job.
- Examining SQL workloads that may run via Hive, Spark, Presto, etc., certain optimization methodologies were determined. First, two (2) options to search the space of configuration values were examined: iterative execution and model-based execution. The optimization function for both methodologies can be defined as:
-
- Looking at the equation above, n is the number of containers launched, βi is the container memory for the ith container launched, and Δiis the execution time for the ith container launched for job. The product is a proxy for the cost of running a container. The sum is a proxy for the cost of running a query. This represents both memory and CPU resources consumed, and may correlate to both dollar cost and throughput of a big data stack.
- Next, parameters that control parallelism and memory per task were examined, as these classes of parameters may have the biggest impact on SQL workloads. For example, such parameters for Map Reduce may include, but are not limited to mapreduce.map.memory .mb, mapreduce.reduce.memory.mb. mapreduce.input.fileinputformat.split.maxsize, mapreduce.input.fileinputformat.split.minsize, and/or hive.exec.reducers.bytes.per.reducer.
- Parameters for Spark may include, but are not limited to spark.executor.cores, spark.executor.memory, spark.hadoop.mapreduce.input.fileinputformat.split.maxsize, spark.hadoop.mapreduce.input.fileinputformat.split.minsize, spark.sql.shuffle.partitions, and/or spark.sql.autoBroadcastJoinThreshold.
- Using an iterative execution approach, Hive queries with various configuration parameters were run, and the most optimal configuration was selected. Strategies to reduce the parameter space may include, but are not limited to: (i) parameter reduction—focus on a small(er) set of configuration parameters: (ii) discretization—by discretizing each parameter fewer values are tried, rather than all possible values for each parameter: (iii) reduce search range—although parameter may be selected from a large range of values, a search may be limited to an expected good range for each dimension using heuristics; and/or (iv) assume dimension independence—ignore dependence of parameters to prevent parameter space explosion due to correlation. Note that as briefly mentioned above, for repetitive queries, each iteration of the query may be used to improve performance or cost.
- An iterative algorithm that searched the space of configuration values based on these constraints was then performed. With reference to
FIG. 2 , aniterative algorithm 20 that searched the space of all configuration values based on these constraints.FIG. 2 illustrates steps to search optimal values for two parameters.Graphical representation 210 shows the search space for two parameters. Graphical representation 220 shows discrete values being explored in each axis. Graphical representation 230 illustrates the search space being restricted based on domain knowledge. Subsequently, graphical representation 240 shows an algorithm reiterating itself through each parameter, choosing an optimal point, and then moving to the next parameter. - An algorithm in accordance with some embodiments of the present invention was then run in an attempt to optimize three (3) different Hive queries. As illustrated in
FIG. 3 , reduction of costs using the auto-tuning processes of the present invention returned cost reductions ranging from 15% to 75%.FIG. 3 illustrates a chart 30, showing data for three different queries (311, 312, 333), depicted against a vertical axis 320 showing percentage of cost reduction. Thefirst query 311 showed a cost reduction of approximately 65%, shown at 331. Thesecond query 312 showed a cost reduction of approximately 75%, shown at 332. Thethird query 313 showed a cost reduction of approximately 15%. Again, each auto-tuned result showed significant improvement and savings over settings chosen by database administrators. However, despite cost savings, this iterative approach was run considered to be practical due to increased cost of running the approach. The high cost was a factor of both the number of queries and iterations, and the use of shadow clusters and tables in ETL queries that required multiple reviews, increasing man-hour associated costs. - Since iterative execution is impractical at scale, a model-based approach to eliminate execution of queries was examined. An execution model was created that substantially replicated an execution engine, based on a reduced set of parameters. In accordance with some embodiments of the present invention, this model may also utilize various statistics about data sizes and selectivities of various operators as input. Such statistics may be obtained by collecting metrics from a previous run, or by using statistics from a database catalog. Metrics from a previous run may be suitable for ETL or reporting queries, while statistics from a database catalog may be suitable for ad-hoc queries. This model outputs the result of the optimization function as described above.
- To quantify the prediction error by the model, four (4) queries were processed. The
graph 40 atFIG. 4 illustrates the benefit predicted by our model and the actual observed benefit for these queries. Each query (411, 412, 413, 414) is set forth along the x axis, while they axis 420 shows the percentage of cost reduction. It can be seen that the first query 411 showed a predicted cost reduction (431A) of approximately 82%, while an observed cost reduction (431B) was approximately 76%. The second query 412 showed a predicted cost reduction (432A) of approximately 80%, while an observed cost reduction (432B) was approximately 82%. The third query 413 showed a predicted cost reduction (433A) of approximately 72%, while an observed cost reduction (433B) was approximately 70%. The fourth query 414 showed a predicted cost reduction (434A) of approximately 80%, while an observed cost reduction (434B) was approximately 78%. It can be seen that, in general, the actual savings closely match the predicted savings indicating that the model is sufficiently accurate. - Some insights and heuristics to optimize SQL workloads were gained. Some are discussed briefly below.
- Insight 1: Container Shape Should Match Instance Family Shape. Yarn allocates containers on two dimensions—memory and virtual central processing unit (vcpu). Each container is given 1 vcpu and some memory. The memory/vcpu of the containers should match the memory/vcpu ratio of the machine type. Otherwise resources may be unnecessarily wasted.
- Insight 2: Avoid Spills in Tasks. Spills are expensive because each spill leads to an extra write and read of all the data. Spills should be avoided at all costs. In general, spills may be avoided by providing adequate memory to each task.
- Insight 3: Decrease Memory Per Task to Choose Cheaper Instance Type. (On cloud platforms, machines with higher memory/cpu are generally more expensive for the same CPU type. Decrease the memory per task and consequently increase parallelism to choose a cheaper instance type. As long as tasks do not spill, the total work done in terms of IO, CPU and network traffic is independent of the parallelism factor of the tasks. For example, the total data read and processed will be the same if the number of mappers is 100 or 1000. If a job can be tuned to avoid spills on a cheaper instance with same compute but lesser memory than original instance, then it is generally a good idea to move to cheaper instance for saving cost without any performance degradation.
- Insight 4: Beware of Secondary Effects of High Parallelism. Parallelism cannot be increased indefinitely. There are secondary effects of increasing the number of tasks. For example every task has to pay the cost of Java virtual machine (JVM) start if applicable. There may also be an increase in number of communication channels. Accordingly, parallelism should be not be set so high that secondary effects drown the increase in performance. This limit is specific to a workload or query and cluster configuration and can be determined algorithmically.
- Insight 5: In Spark Environments, Prefer Fat Executors. This insight is specific to Spark, where there is an additional parameter of cores per executor. Given a certain number of cores per machine, there may be a choice of either running many executors with fewer cores per executor (thin executors), or fewer executors with more cores per executor (fat executors). It has been observed that fat executors generally provide better performance, due to better memory utilization across cores, reduced number of replicas of broadcast tables, and lesser overheads due to more tasks running in the same executor process.
- It will be understood that the specific embodiments of the present invention shown and described herein are exemplary only. Numerous variations, changes, substitutions and equivalents will now occur to those skilled in the art without departing from the spirit and scope of the invention. Accordingly, it is intended that all subject matter described herein and shown in the accompanying drawings be regarded as illustrative only, and not in a limiting sense.
Claims (21)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US16/102,852 US11228489B2 (en) | 2018-01-23 | 2018-08-14 | System and methods for auto-tuning big data workloads on cloud platforms |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US201862620543P | 2018-01-23 | 2018-01-23 | |
US16/102,852 US11228489B2 (en) | 2018-01-23 | 2018-08-14 | System and methods for auto-tuning big data workloads on cloud platforms |
Publications (2)
Publication Number | Publication Date |
---|---|
US20190229992A1 true US20190229992A1 (en) | 2019-07-25 |
US11228489B2 US11228489B2 (en) | 2022-01-18 |
Family
ID=67300249
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US16/102,852 Active US11228489B2 (en) | 2018-01-23 | 2018-08-14 | System and methods for auto-tuning big data workloads on cloud platforms |
Country Status (1)
Country | Link |
---|---|
US (1) | US11228489B2 (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112181603A (en) * | 2020-10-26 | 2021-01-05 | 浪潮云信息技术股份公司 | Controllable vertical automatic retractor implementation method and system in cloud environment |
US20210065130A1 (en) * | 2019-08-28 | 2021-03-04 | Yu-Ting Lee | Personal information platform system incorporating big data and artificial intelligence |
CN112445776A (en) * | 2020-11-20 | 2021-03-05 | 北京易观智库网络科技有限公司 | Presto-based dynamic barrel dividing method, system, equipment and readable storage medium |
US10958521B2 (en) * | 2019-07-19 | 2021-03-23 | Oracle International Corporation | Method and apparatus for configuring a cloud storage software appliance |
US20210365302A1 (en) * | 2020-05-19 | 2021-11-25 | Hewlett Packard Enterprise Development Lp | Adaptive and distributed tuning system and method |
WO2022261959A1 (en) * | 2021-06-18 | 2022-12-22 | 深圳市大疆创新科技有限公司 | Leveling method and gimbal |
US11561843B2 (en) | 2020-06-11 | 2023-01-24 | Red Hat, Inc. | Automated performance tuning using workload profiling in a distributed computing environment |
US20230367783A1 (en) * | 2021-03-30 | 2023-11-16 | Jio Platforms Limited | System and method of data ingestion and processing framework |
US11841878B2 (en) | 2021-08-30 | 2023-12-12 | International Business Machines Corporation | Automatic vertical partitioning of fact tables in a distributed query engine |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130232254A1 (en) * | 2012-03-02 | 2013-09-05 | Computenext Inc. | Cloud resource utilization management |
US20170337138A1 (en) * | 2016-05-18 | 2017-11-23 | International Business Machines Corporation | Dynamic cache management for in-memory data analytic platforms |
Family Cites Families (74)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5603001A (en) | 1994-05-09 | 1997-02-11 | Kabushiki Kaisha Toshiba | Semiconductor disk system having a plurality of flash memories |
US5907675A (en) | 1995-03-22 | 1999-05-25 | Sun Microsystems, Inc. | Methods and apparatus for managing deactivation and shutdown of a server |
US6192391B1 (en) | 1997-05-30 | 2001-02-20 | Nec Corporation | Process stop method and apparatus for a distributed memory multi-processor system |
US6460027B1 (en) | 1998-09-14 | 2002-10-01 | International Business Machines Corporation | Automatic recognition and rerouting of queries for optimal performance |
US6847962B1 (en) | 1999-05-20 | 2005-01-25 | International Business Machines Corporation | Analyzing, optimizing and rewriting queries using matching and compensation between query and automatic summary tables |
US6789213B2 (en) | 2000-01-10 | 2004-09-07 | Sun Microsystems, Inc. | Controlled take over of services by remaining nodes of clustered computing system |
US6601062B1 (en) | 2000-06-27 | 2003-07-29 | Ncr Corporation | Active caching for multi-dimensional data sets in relational database management system |
US6735587B2 (en) | 2000-07-28 | 2004-05-11 | International Business Machines Corporation | Maintaining pre-computed aggregate views incrementally in the presence of non-minimal changes |
US6918051B2 (en) | 2001-04-06 | 2005-07-12 | International Business Machines Corporation | Node shutdown in clustered computer system |
US20020157113A1 (en) | 2001-04-20 | 2002-10-24 | Fred Allegrezza | System and method for retrieving and storing multimedia data |
US20030005350A1 (en) | 2001-06-29 | 2003-01-02 | Maarten Koning | Failover management system |
US6901410B2 (en) | 2001-09-10 | 2005-05-31 | Marron Pedro Jose | LDAP-based distributed cache technology for XML |
JP4562115B2 (en) | 2003-06-11 | 2010-10-13 | ソニー株式会社 | File generating apparatus and method, program, and recording medium |
US8886617B2 (en) | 2004-02-20 | 2014-11-11 | Informatica Corporation | Query-based searching using a virtual table |
US20050222996A1 (en) | 2004-03-30 | 2005-10-06 | Oracle International Corporation | Managing event-condition-action rules in a database system |
JP4574230B2 (en) | 2004-05-27 | 2010-11-04 | 株式会社日立製作所 | Database management method and database system |
US7475204B2 (en) | 2004-11-24 | 2009-01-06 | International Business Machines Corporation | Automatically managing the state of replicated data of a computing environment |
US8996482B1 (en) | 2006-02-10 | 2015-03-31 | Amazon Technologies, Inc. | Distributed system and method for replicated storage of structured data records |
US7848261B2 (en) | 2006-02-17 | 2010-12-07 | Isilon Systems, Inc. | Systems and methods for providing a quiescing protocol |
JP2008123357A (en) | 2006-11-14 | 2008-05-29 | Honda Motor Co Ltd | Parallel computer system, parallel computing method, and program for parallel computer |
US7844853B2 (en) | 2007-08-07 | 2010-11-30 | International Business Machines Corporation | Methods and apparatus for restoring a node state |
US7882087B2 (en) | 2008-01-15 | 2011-02-01 | At&T Intellectual Property I, L.P. | Complex dependencies for efficient data warehouse updates |
US20090222418A1 (en) | 2008-02-29 | 2009-09-03 | Layman Timothy B | Systems and methods for dynamic content presentation |
US8156421B2 (en) | 2008-06-30 | 2012-04-10 | Yahoo! Inc. | Analysis of database performance reports for graphical presentation of summary results |
US20100153482A1 (en) | 2008-12-10 | 2010-06-17 | Full Armor Corporation | Cloud-Based Automation of Resources |
US9178935B2 (en) | 2009-03-05 | 2015-11-03 | Paypal, Inc. | Distributed steam processing |
US8296419B1 (en) | 2009-03-31 | 2012-10-23 | Amazon Technologies, Inc. | Dynamically modifying a cluster of computing nodes used for distributed execution of a program |
US8909863B2 (en) | 2009-11-16 | 2014-12-09 | Microsoft Corporation | Cache for storage and/or retrieval of application information |
US8424052B2 (en) | 2009-12-18 | 2013-04-16 | Samir ABED | Systems and methods for automated extraction of closed captions in real time or near real-time and tagging of streaming data for advertisements |
US8452932B2 (en) * | 2010-01-06 | 2013-05-28 | Storsimple, Inc. | System and method for efficiently creating off-site data volume back-ups |
CN102209087B (en) | 2010-03-31 | 2014-07-09 | 国际商业机器公司 | Method and system for MapReduce data transmission in data center having SAN |
US8260840B1 (en) | 2010-06-28 | 2012-09-04 | Amazon Technologies, Inc. | Dynamic scaling of a cluster of computing nodes used for distributed execution of a program |
US20120047339A1 (en) | 2010-08-20 | 2012-02-23 | Dan Decasper | Redundant array of independent clouds |
US20120102291A1 (en) | 2010-10-26 | 2012-04-26 | Dell Products, Lp | System and Method for Storage Allocation in a Cloud Environment |
WO2012063294A1 (en) | 2010-11-12 | 2012-05-18 | 株式会社日立製作所 | Computer system |
US8560887B2 (en) | 2010-12-09 | 2013-10-15 | International Business Machines Corporation | Adding scalability and fault tolerance to generic finite state machine frameworks for use in automated incident management of cloud computing infrastructures |
US20120215763A1 (en) | 2011-02-18 | 2012-08-23 | Microsoft Corporation | Dynamic distributed query execution over heterogeneous sources |
US20120304192A1 (en) | 2011-05-27 | 2012-11-29 | International Business Machines | Lifeline-based global load balancing |
IL216056B (en) | 2011-10-31 | 2018-04-30 | Verint Systems Ltd | Combined database system and method |
US9049746B2 (en) | 2011-11-08 | 2015-06-02 | Blackberry Limited | Management of multi-mode wireless devices |
WO2013070873A1 (en) | 2011-11-10 | 2013-05-16 | Treasure Data, Inc. | System and method for operating a big-data platform |
US9122535B2 (en) | 2011-11-22 | 2015-09-01 | Netapp, Inc. | Optimizing distributed data analytics for shared storage |
US9170849B2 (en) | 2012-01-09 | 2015-10-27 | Microsoft Technology Licensing, Llc | Migration of task to different pool of resources based on task retry count during task lease |
US8732291B2 (en) | 2012-01-13 | 2014-05-20 | Accenture Global Services Limited | Performance interference model for managing consolidated workloads in QOS-aware clouds |
US9172608B2 (en) | 2012-02-07 | 2015-10-27 | Cloudera, Inc. | Centralized configuration and monitoring of a distributed computing cluster |
US9268590B2 (en) | 2012-02-29 | 2016-02-23 | Vmware, Inc. | Provisioning a cluster of distributed computing platform based on placement strategy |
US9483785B1 (en) | 2012-05-07 | 2016-11-01 | Amazon Technologies, Inc. | Utilizing excess resource capacity for transcoding media |
WO2013169249A1 (en) | 2012-05-09 | 2013-11-14 | Intel Corporation | Scheduling tasks among processor cores |
US9531607B1 (en) | 2012-06-20 | 2016-12-27 | Amazon Technologies, Inc. | Resource manager |
US9223845B2 (en) | 2012-08-01 | 2015-12-29 | Netapp Inc. | Mobile hadoop clusters |
US8935495B2 (en) | 2012-08-21 | 2015-01-13 | International Business Machines Corporation | Storage management in a virtual environment |
US20140059552A1 (en) | 2012-08-24 | 2014-02-27 | International Business Machines Corporation | Transparent efficiency for in-memory execution of map reduce job sequences |
JP2014044677A (en) | 2012-08-28 | 2014-03-13 | Fujitsu Ltd | Transmission control program, communication node, and transmission control method |
US20150222705A1 (en) | 2012-09-06 | 2015-08-06 | Pi-Coral, Inc. | Large-scale data storage and delivery system |
US9323767B2 (en) | 2012-10-01 | 2016-04-26 | Longsand Limited | Performance and scalability in an intelligent data operating layer system |
US10394611B2 (en) | 2012-11-26 | 2019-08-27 | Amazon Technologies, Inc. | Scaling computing clusters in a distributed computing system |
US9385915B2 (en) | 2012-11-30 | 2016-07-05 | Netapp, Inc. | Dynamic caching technique for adaptively controlling data block copies in a distributed data processing system |
US9645859B1 (en) | 2012-12-18 | 2017-05-09 | Veritas Technologies Llc | Performing I/O quiesce and drain operations in multi-node distributed systems |
KR101540631B1 (en) | 2012-12-28 | 2015-07-30 | 삼성에스디에스 주식회사 | System, method and recording medium recording the program thereof for dynamic expansion of the virtual cluster |
US9451013B1 (en) | 2013-01-02 | 2016-09-20 | Amazon Technologies, Inc. | Providing instance availability information |
US9081826B2 (en) * | 2013-01-07 | 2015-07-14 | Facebook, Inc. | System and method for distributed database query engines |
WO2014138745A2 (en) | 2013-03-08 | 2014-09-12 | Curalate, Inc. | Method and system for clustering and classifying online visual information |
EP2973051A4 (en) | 2013-03-15 | 2016-11-16 | Amazon Tech Inc | Scalable analysis platform for semi-structured data |
US10325032B2 (en) | 2014-02-19 | 2019-06-18 | Snowflake Inc. | Resource provisioning systems and methods |
US9880826B2 (en) | 2014-02-25 | 2018-01-30 | Red Hat, Inc. | Installing of application resources in a multi-tenant platform-as-a-service (PaS) system |
US9852138B2 (en) | 2014-06-30 | 2017-12-26 | EMC IP Holding Company LLC | Content fabric for a distributed file system |
US20160224638A1 (en) | 2014-08-22 | 2016-08-04 | Nexenta Systems, Inc. | Parallel and transparent technique for retrieving original content that is restructured in a distributed object storage system |
US9894119B2 (en) | 2014-08-29 | 2018-02-13 | Box, Inc. | Configurable metadata-based automation and content classification architecture for cloud-based collaboration platforms |
US10069693B1 (en) | 2014-12-11 | 2018-09-04 | Amazon Technologies, Inc. | Distributed resource allocation |
US20160179581A1 (en) | 2014-12-19 | 2016-06-23 | Netapp, Inc. | Content-aware task assignment in distributed computing systems using de-duplicating cache |
US10025823B2 (en) | 2015-05-29 | 2018-07-17 | Oracle International Corporation | Techniques for evaluating query predicates during in-memory table scans |
US9779031B2 (en) | 2015-06-17 | 2017-10-03 | International Business Machines Corporation | Caching policies for selection and replacement of objects |
US9860569B1 (en) | 2015-10-05 | 2018-01-02 | Amazon Technologies, Inc. | Video file processing |
US10713073B2 (en) * | 2016-12-02 | 2020-07-14 | Microsoft Technology Licensing, Llc | Systems and methods for identifying cloud configurations |
-
2018
- 2018-08-14 US US16/102,852 patent/US11228489B2/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130232254A1 (en) * | 2012-03-02 | 2013-09-05 | Computenext Inc. | Cloud resource utilization management |
US20170337138A1 (en) * | 2016-05-18 | 2017-11-23 | International Business Machines Corporation | Dynamic cache management for in-memory data analytic platforms |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10958521B2 (en) * | 2019-07-19 | 2021-03-23 | Oracle International Corporation | Method and apparatus for configuring a cloud storage software appliance |
US11362893B2 (en) | 2019-07-19 | 2022-06-14 | Oracle International Corporation | Method and apparatus for configuring a cloud storage software appliance |
US20210065130A1 (en) * | 2019-08-28 | 2021-03-04 | Yu-Ting Lee | Personal information platform system incorporating big data and artificial intelligence |
US20210365302A1 (en) * | 2020-05-19 | 2021-11-25 | Hewlett Packard Enterprise Development Lp | Adaptive and distributed tuning system and method |
US11561843B2 (en) | 2020-06-11 | 2023-01-24 | Red Hat, Inc. | Automated performance tuning using workload profiling in a distributed computing environment |
CN112181603A (en) * | 2020-10-26 | 2021-01-05 | 浪潮云信息技术股份公司 | Controllable vertical automatic retractor implementation method and system in cloud environment |
CN112445776A (en) * | 2020-11-20 | 2021-03-05 | 北京易观智库网络科技有限公司 | Presto-based dynamic barrel dividing method, system, equipment and readable storage medium |
US20230367783A1 (en) * | 2021-03-30 | 2023-11-16 | Jio Platforms Limited | System and method of data ingestion and processing framework |
WO2022261959A1 (en) * | 2021-06-18 | 2022-12-22 | 深圳市大疆创新科技有限公司 | Leveling method and gimbal |
US11841878B2 (en) | 2021-08-30 | 2023-12-12 | International Business Machines Corporation | Automatic vertical partitioning of fact tables in a distributed query engine |
Also Published As
Publication number | Publication date |
---|---|
US11228489B2 (en) | 2022-01-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11228489B2 (en) | System and methods for auto-tuning big data workloads on cloud platforms | |
US12010167B2 (en) | Automated server workload management using machine learning | |
US10983895B2 (en) | System and method for data application performance management | |
Bilal et al. | Towards automatic parameter tuning of stream processing systems | |
Gautam et al. | A survey on job scheduling algorithms in big data processing | |
US9043787B2 (en) | System and method for automated assignment of virtual machines and physical machines to hosts | |
Sidhanta et al. | OptEx: A deadline-aware cost optimization model for spark | |
US20180081629A1 (en) | Method and apparatus for providing ordered sets of arbitrary percentile estimates for varying timespans | |
US20140019965A1 (en) | System and method for automated assignment of virtual machines and physical machines to hosts with right-sizing | |
US20120198466A1 (en) | Determining an allocation of resources for a job | |
Taft et al. | P-store: An elastic database system with predictive provisioning | |
Vakilinia et al. | Analysis and optimization of big-data stream processing | |
EP2671152A1 (en) | Estimating a performance characteristic of a job using a performance model | |
JP2016042284A (en) | Parallel computer system, management device, method for controlling parallel computer system, and management device control program | |
US20230004419A1 (en) | Resource Migration Method and System, and Device | |
Sellami et al. | Clustering-based data placement in cloud computing: a predictive approach | |
US11675515B2 (en) | Intelligent partitioning engine for cluster computing | |
EP3091449B1 (en) | Operating a database system | |
US20170235608A1 (en) | Automatic response to inefficient jobs in data processing clusters | |
Li et al. | Data allocation in scalable distributed database systems based on time series forecasting | |
Nascimento et al. | Data quality monitoring of cloud databases based on data quality SLAs | |
Nagarajan et al. | Malleable scheduling for flows of jobs and applications to MapReduce | |
Chandra et al. | An Efficient Framework for Load Balancing using MapReduce Algorithm for Bigdata | |
Giannakouris et al. | Cost-Effective, Workload-Adaptive Migration of Big Data Applications to the Cloud | |
Maghawry et al. | Enhancing query optimization technique by conditional merging over cloud computing |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
FEPP | Fee payment procedure |
Free format text: ENTITY STATUS SET TO UNDISCOUNTED (ORIGINAL EVENT CODE: BIG.); ENTITY STATUS OF PATENT OWNER: SMALL ENTITY |
|
FEPP | Fee payment procedure |
Free format text: ENTITY STATUS SET TO SMALL (ORIGINAL EVENT CODE: SMAL); ENTITY STATUS OF PATENT OWNER: SMALL ENTITY |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: NON FINAL ACTION MAILED |
|
AS | Assignment |
Owner name: QUBOLE INC., CALIFORNIA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:MARGOOR, AMOGH;VENKATESH, RAJAT;REEL/FRAME:052307/0960 Effective date: 20200330 |
|
AS | Assignment |
Owner name: JEFFERIES FINANCE LLC, NEW YORK Free format text: FIRST LIEN SECURITY AGREEMENT;ASSIGNOR:QUBOLE INC.;REEL/FRAME:054498/0115 Effective date: 20201120 Owner name: JEFFERIES FINANCE LLC, NEW YORK Free format text: SECOND LIEN SECURITY AGREEMENT;ASSIGNOR:QUBOLE INC.;REEL/FRAME:054498/0130 Effective date: 20201120 |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: NON FINAL ACTION MAILED |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: RESPONSE TO NON-FINAL OFFICE ACTION ENTERED AND FORWARDED TO EXAMINER |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: EX PARTE QUAYLE ACTION MAILED |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: RESPONSE TO EX PARTE QUAYLE ACTION ENTERED AND FORWARDED TO EXAMINER |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: NOTICE OF ALLOWANCE MAILED -- APPLICATION RECEIVED IN OFFICE OF PUBLICATIONS |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: PUBLICATIONS -- ISSUE FEE PAYMENT RECEIVED |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: PUBLICATIONS -- ISSUE FEE PAYMENT VERIFIED |
|
STCF | Information on status: patent grant |
Free format text: PATENTED CASE |